Group actions and learning for a family of automata (Q1109570)

A study of the dynamics of a family of artificial neural networks (ANN) is presented based on a new method, that of the action of the group of isometries of the state space (hypercube). This group is generated by two kinds of isometries: permutation of two coordinates of a state and sign invertion of one coordinate. After a short review of the Hopfield initial model for content adressable memory there are presented some of its generalizations and several of learning algorithms, i.e., determining interconnection weights knowing the stable points of ANN. In the main section of the paper the new method is presented and it is shown that most of the learning algorithms are invariant under the action of defined isometries. Also it is proved on what conditions this invariance takes place, and what is the general structure of the space of states of an ANN. In the next section a simple application of a theorem proved in the previous section, about the structure of the space of stable points, is presented. Results are derived about an upper bound on the number of stable states for a particular ANN having as learning algorithm the sum of outer products of degree 2. The last section on concluding remarks presents some of possible further work in the field and a proposal for some practical applications. The paper ends with an appendix in which several formal properties are listed that are relevant for learning rules.